Maintain the documentation¶

Write for someone using or extending Repo2RLEnv. Describe the implemented behavior, its limits and a runnable example. Put campaign diaries, provider receipts, budget approvals and one-off input selections in ignored workspace/ directories.

Each pipeline guide should explain its inputs, show one stage diagram, map model calls to their inputs and outputs, and describe verification and bounded repair. Link to canonical prompts, upstream credits, measured economics and the dataset. Keep shared execution and quality contracts in their common guides and RFCs.

Build and preview¶

python -m pip install -r requirements-docs.txt
mkdocs serve
mkdocs build --strict

No model keys, cloud accounts, private campaign directories or installed research repositories are needed. The build reads owned source files as text; it does not execute task code.

The MkDocs pre-build hook generates docs/pipelines/prompts/ from canonical templates and request-assembly code. That directory and site/ are ignored. Edit the canonical prompt or the generator, then rebuild; do not commit copies of generated pages. The detailed reference remains available in the built site, while each guide summarizes the effective model request and links to the source for GitHub readers.

Update measured results¶

Review the evidence and update docs/data/pipelines.json, then run:

python docs/_tools/generate_metrics.py
python docs/_tools/generate_metrics.py --check

Keep only sanitized aggregate measurements and public dataset revisions in this file. Count retries under their candidate identity, separate retained tasks, and keep generation, quality evaluation and solver outcomes distinct. Costs must name their sample and include failures; unavailable compute or yield is not zero. Keep raw evidence locally or with an appropriate dataset/release artifact.

Historical native-pipeline records live in the same file under native_history. Preserve their evidence scope: a cached inventory, a generation-time verification stamp and a Harbor oracle gate are different observations. Do not carry a gate from an older cohort onto a newer dataset. The old synthesis counters are run-cumulative; sum one final counter per run, not every exported task's counter. Update native results when changing those measurements, and retain source hashes or pinned public manifests.

Review changes¶

Check the stage diagram against the implementation, validate example configs and build the site from a clean checkout. CI checks the generated prompt references, metrics tables, links and site build. Inspect changed diagrams in a browser. Preserve upstream notices and source pins when condensing a guide or RFC.